PiedPiper (jamesdgwise@gmail.com) on 4/13/11 wrote:
---------------------------
>Another good article, and i definitely like the idea of more regular smaller, and
>as a result simpler/easier to understand,
>articles.
>
>A good if somewhat basic analysis of modern GPU's.
>I'm interested as to weather or not you did any testing / theorizing that took
>into account memory bandwidth, prior to ignoring that >factor?

Absolutely. I decided that I didn't have the time or word count to really deal with it adequately. Bandwidth matters, but one step at a time.

>I cant help but think *purely guessing though* that if you added a scaling factor
>that accounted for differences in memory bandwidth, that your results might be even
>more accurate, especially when generalizing from the >laptop to the desktop version of the cards.

That's definitely true, and the models I use for myself include memory bandwidth. But as I said in the article - it's more complicated, harder to visualize and really just beyond the scope of what I wanted to do.

>EG. Memory bandwidth / number of processing cores
>( or maybe establish a baseline memory Bandwidth for each >model of GPU, and then compare against that....
>
>Also interested as to why you chose notebookcheck.net as >your data source? Why
>not use Anand/Hardocp/techreport and compare desktop GPU's?

Of the three you mentioned, I only really trust data from Scott @ TR. Unfortunately, Scott doesn't use 3DMark, and I didn't want to use a real game (for a variety of reasons which might become clear later). Moreover, none of them put their data into a form that is readily usable.

>Oh and i would have been interested to see results from the new VLIW4 ati cards,
>even if the are outliers and represent a different >architecture.

There were only 2 data points, which is the bare minimum needed for any sort of analysis. I'm just not very comfortable doing any sort of modeling with that little data. Perhaps in the future though...

David